Welcome Beam: a 501B open model you can own, customize, and scale
huggingface ·
Reflection has released Beam, a new 501-billion parameter open-source language model available under the Apache 2.0 license on Hugging Face. Designed primarily for coding and agentic tasks, Beam utilizes a Mixture-of-Experts architecture with 23 billion active parameters per token while supporting a massive one-million-token context window. The model features specialized attention mechanisms and residual stream handling to maintain performance over long contexts, alongside native support for tool calling and structured reasoning. Users can deploy Beam using day-zero compatible libraries such as Transformers, vLLM, and SGLang, with checkpoints optimized for various hardware including NVIDIA Blackwell, Hopper, and Ampere GPUs. The article details benchmark results across coding, reasoning, and general capability tests, demonstrating strong performance in agentic workflows. Additionally, it provides comprehensive guides for post-training techniques like supervised fine-tuning with TRL and reinforcement learning via Miles. To encourage community adoption, Reflection is offering $10,000 in Hub credits for innovative projects built with Beam. Several demos illustrate practical applications, including building native macOS applications and scaling multi-agent programming tasks.
huggingface